Machine learning can predict setting behavior and strength evolution of hydrating cement systems

Abstract

AbstractSetting and strength development of ordinary Portland cement (OPC) binders involves multiple interacting chemical reactions, resulting in the formation of a solid microstructure. A long‐standing yet elusive goal has been to establish a basis for the prediction of the properties and performance of concrete using knowledge of the chemical and physical attributes of its components—PC, sand, stone, water, and chemical admixtures—together with the environmental conditions under which they react. Machine learning (ML) provides a data‐driven basis for the estimation of properties, and has recently been applied to estimate the 28 days (compressive) strength of concrete from knowledge of its mixture proportions (Young et al, Cem Concr Res, 2019, 115:379). Building on this success, the current work uses a diverse dataset of ASTM C150 cements, the chemical composition and other attributes of which have been measured. ML estimators were trained with this dataset to estimate both paste setting time and mortar strength development. The ML estimation errors are typically similar to the measurement repeatability of the relevant ASTM test methods, and are thus able to account for the influence of binder composition and fineness. This creates new opportunities to apply data intensive methods to optimize concrete formulations under multiple constraints of cost, CO2 impact, and performance attributes.